Evaluating Explainable AI Methods for Geoscientific Regression: Insights from Applications and the Lorenz-63 System

XAI methods designed for classification tasks may not work well for regression problems, which are more common in Earth system science and geoscientific applications, suggesting that existing explainability tools need adaptation for these domains.

Even with explainable AI techniques, complex AI models can fail in unexpected ways, as demonstrated through testing on the Lorenz 1963 system—a physically meaningful benchmark that revealed limitations and failure modes in general-purpose XAI methods when applied to chaotic systems.

Different stakeholders in operational Earth system science have different needs for AI explainability, indicating that a one-size-fits-all approach to XAI is insufficient and methodological choices should be linked to specific user groups and their decision-making requirements.

Building trust in AI forecasting tools requires understanding both what predictions are made and why, but the "black box" nature of modern AI systems creates challenges that current XAI approaches have not fully solved, particularly for the regression and forecasting tasks critical to Earth science.

Significant gaps remain in existing XAI methodologies for geoscientific applications, suggesting that developing domain-specific explainability approaches tailored to Earth system science is necessary before these AI systems can be reliably deployed in operational settings.